Modern software development faces unprecedented complexity. Multi-architecture deployments, regulatory requirements, and long-lived systems demand more than isolated coding tools. Present day AI-assisted coding state-of-the art is limited to generation of code snippets and refactor functions, Within the AI4SWEng project, the University of Reading, as the scientific and technical leader, has helped address the challenges of the next frontier in LLM-powered developer support for integrated system-level Requirements Analysis, Design, Development and Automated Testing to deliver robust self-evolving systems.
The AI4SWEng framework supports Software Engineering as a controlled, end-to-end integrative process, rather than a sequence of disjointed tasks. Its goal is not to replace developers, but to provide a structured environment where AI capabilities can be harnessed to support efficient solution development safely, traceably, and at scale.
For AI4SWEng to support the future Software-AI Engineers to efficiently deliver fully conformant, reliable and robust solutions, the design and development process must be underpinned by a methodologically-guided analysis base that systematically supports end-to-end and lifecycle compliance-by-design. This must be rooted in an ontological commitment to dynamically adaptive requirements re-prioritisation supported by use-contexts-specific indexicality of each requirement to enable deep track-and-trace of responsibility and responsiveness of every component, data and control flow at every level of system design and behaviour. To this end AI4SWEng has deployed the UI-REF Methodology [1,2] to set the foundations for accountability-reliability engineering and the compliance assurance of functionally and socio-ethically safe, secure and responsively self-adaptive systems. UI-REF is the only methodology ever devised with context-specific requirements elicitation and usability probing methods and analysis base that are consistent with long-established psycho-cognitive theories governing humans’ situated patterns-of-relating to self and others including human-machine relationship evolution based on the principles of Human Judgement and Decision Making Theory [3,4] such as recency-saliency and thus the Memorability of Experiences and their Recall, particularly of Pleasure-and-Pain points in user experiences with machines. In this way UI-REF supports the dynamic re-adaptation of system performance responsive to evolving user preferences and real-world data changes (model drifts). It is essentially a holistic methodology for Design-by-Contract including socio-ethical and regulatory compliance by design. In this way, it ensures the design, development and deployment of trustworthy solutions with trade-offs dynamically re-optimised responsive to the situated use-contexts per stakeholder preferences.
At the core of this framework lies the AI-SysDev Platform, a modular, DevOps-oriented architecture led by UREAD, with key contributions from Partner Innova Integra (INNO). This platform integrates AI-driven development, testing, governance, and man-in-the-loop, and serves as the primary input and foundation for all subsequent development tasks within the AI4SWEng project.
From Toolchains to Coordinated Systems
Traditional software environments evolve incrementally: requirements tools, version control, CI/CD pipelines, testing frameworks, and security checks are added over time, often with minimal integration. As complexity increases, coordination becomes manual, error-prone, and difficult to audit.
AI4SWEng introduces a coordination-first architecture, where all development activities are orchestrated through a shared control and integration layer. The platform follows an Enterprise Service Bus (ESB)-inspired design, enforcing clear interfaces, shared context, and controlled interaction between components. This enables individual tools and AI models to evolve independently while preserving system-wide conformance which is a necessity in environment where AI technologies and regulatory requirements chang rapidly.
Framework Architectural Specification for the AI-SysDev Platform
As mandated by our methodologically-guided approach (UI-REF), the AI4SWEng Framework Architecture Specification was based on detailed user-centred requirements analysis including use-context-specific conflicts resolution and overall prioritisation of functional and non-functional system performance requirements to support the 13 Key Innovation Outputs (KIOs) of the project. Accordingly, the AI-SysDev Platform architecture was envisioned as comprising seven building blocks that support service layers, each responsible for a distinct phase of the software development lifecycle – together, they create a unified, auditable environment; as set out below and illustrated in Figures 1, 2 and 3 as follows:
- GUI and Execution
This layer acts as the primary interface between humans and the system. Users submit requirements, inspect generated artefacts, execute workflows, and intervene when necessary. This block also hosts the Developer Training Tool (KIO13), ensuring developers remain skilled and in control.
- Orchestration and Integration
This layer is powered by KIO1, this block coordinates all other KIOs. As such, KIO1 manages workflow sequencing, task decomposition, shared state propagation, and governance enforcement. Centralised orchestration ensures the platform behaves as a cohesive system, not a collection of disconnected tools.
- Requirements Engineering
This layer transforms stakeholder intent, as natural language expressions, into structured, machine-interpretable specifications. This requirements formalisation stage reduces ambiguity and ensures traceability throughout the downstream activities-from code generation to testing and compliance assurance.
- Core AI Engines
This layer hosts the AI-driven development capabilities. The AI-SysDev Tool (KIO7) generates and refines code, applies bug fixes, and provides architectural reasoning -all under orchestration control.
- Data Management
This layer provides persistent, shared system memory. Knowledge graphs, vector databases, and provenance-aware storage ensure context is preserved and re-used, thus enabling consistent AI reasoning and reproducibility.
- Test Automation and DevOps
This layer integrates CI/CD pipelines, cross-compilation, and automated testing. The Test Automation Tool (KIO11) verifies correctness, robustness, and regression behaviour as part of the workflow rather than as a final step.
- Security and Governance
This layer ensures compliance, security, and ethical standards throughout the lifecycle. Governance is continuous, embedding traceability and auditability into every artefact. Security-privacy is assured through the UI-REF-based Ethical and Regulatory Compliance Assurance Framework (ERCAF) including security-privacy protection by design and countermeasures prioritisation as set out in Deliverable D1.2 of this project.
An AI Engineering Platform Built Around Lifecycle Intelligence
The platform structures AI-assisted software development around a sequence of lifecycle stages, each supported by specialised KIOs. Rather than presenting AI as a single all-purpose tool, the platform decomposes software engineering into explicit, connected stages-from elicited expressions of intended requirements to delivering final deployment-ready solution -while maintaining traceability, governance, and continuous feedback through HITL support.
The lifecycle begins with user-conceptualised requirements. User’s needs as expressed in natural language are transformed into structured, formalised user stories through KIO3, reducing ambiguity early in the process. These formal requirements are then translated into architecture-level design options by KIO4, enabling design decisions to be reasoned about, compared, and evolved before the code is written.
In this way, once the intended output of the desired solution system to be developed is formally established, the platform addresses data readiness. KIO6 validates, cleans, and balances datasets to ensure quality and consistency of data input stream, while KIO5 generates privacy-preserving synthetic data when real data is sensitive, scarce, or incomplete. This ensures that downstream AI driven‑ development and testing are provided with high quality balanced data sets.
The implementation phase is driven by KIO7, which applies fine-tuned large language models and retrieval-augmented generation to produce, refine, and optimise code and design artefacts. Code execution is not treated as a final step: runtime behaviour is analysed through KIO2, which uses reverse execution and dynamic slicing to localise faults and enable automated bug fixing and code refinements informed by execution results which are fed back directly into code re-engineering and re-design, thus closing the loop between development and validation.
Quality, security, and deployability are enforced continuously rather than at the end of the lifecycle. KIO11 integrates AI-powered test automation-combining fuzz testing, symbolic execution, and security validation, directly into DevOps pipelines. For real-world deployment, KIO8 supports cross-compilation and automated re-configuration across heterogeneous hardware platforms, while KIO10 optimises models and software for energy-efficient execution on resource-constrained edge and embedded systems-Smart Cyber-Physical Systems, TinyML [5].
Throughout all lifecycle stages, responsibility and human oversight are embedded by design. KIO9 enforces ethical, privacy, and regulatory constraints using ontology-driven, design-by-contract mechanisms, ensuring compliance is intrinsically built-in rather than retrofitted. KIO13 supports developers through targeted training and structured prompt engineering, enabling informed supervision and effective collaboration with AI systems.
All activities are coordinated through KIO1, the AI Engineering Suite, which orchestrates workflows, maintains shared context, and ensures traceability across artefacts and decisions. Together, the KIOs support an integrative lifecycle support system for Software-AI Engineering in which requirements, data, code, tests, and compliance co-evolve -reducing manual overhead while increasing confidence in AI-assisted software development.
Example Interaction Flow: From Requirements Statements to System Performance Results Review
A typical user interaction with the AI4SWEng platform follows a controlled, end-to-end workflow as shown in Figure 3.
- Translating Requirements Intentions to Formal Requirements Specification
Stakeholders/Users submit their intended requirements as natural-language expressions input to the GUI. KIO1 routes them to the appropriate requirements KIOs for formalisation and traceability. - Coding
Once the requirements are formalised, KIO1 invokes the AI-SysDev Tool (KIO7) to generate or refine the required software artefacts, accordingly, using the context from the data layer. - Testing
Software artefacts are passed to the Test Automation Tool (KIO11), which verifies correctness, robustness, and compliance automatically. - Performance Results Review
Results are presented via the GUI. The Orchestrator Human-in-the Loop (HITLP) intervention logic supports the exercise of human review as required based on rules and governance constraints. Contextual guidance is provided through KIO13 to assist the human actor in understanding and decision-making before proceeding to any revision steps or approval of the resulting software artefact for deployment. Throughout this process, KIO1 maintains system state, enforces governance, and records provenance for full auditability.
From Writing Code to Orchestrating Intelligent Systems
AI4SWEng is set to transform software engineering from manual code production to intelligent system orchestration. By coordinating specialised KIOs under centralised orchestration and human supervision, the platform enables scalable automation without sacrificing control.
Individual KIOs can evolve independently, while system-level behaviour remains predictable, auditable, and compliant. Cryptographic provenance passporting ensures that every artefact-from requirement to deployed binary- remains fully traceable.
The result is faster, more reliable, and accountable software. In this way, Software-AI Engineers are liberated from the tyranny of forests of code snippets to overseeing the design-develop-test process as powerfully supported by the intelligently orchestrated software production system – the blueprint for modern, enterprise-ready software engineering.
Requirements Engineering for AI4SWEng Developers’ Training Programme (AIEngDevTRN)
The University of Reading Team have completed the UI-REF-based structuring of the AIEngDevTRN programme as a set of 12 multimedia Training Companions each focused on one of the 12 Developer Support Services (KIOs) as offered by AI4SWEng tools to facilitate the lifecycle design, development, evaluation and maintenance of Software-AI-Engineering solution stacks.
This includes the model-driven design and development of Smart Cyber-Physical Systems (S-CPS) as in TinyML which is the focus of joint innovation efforts with our Partner AI4SEC.
Requirements Engineering for Use-Context-Specific Battery Lifecycle Modelling
A notable use-case for the AI4SWEng TinyML is the Prediction of Remaining Useful Life (RUL) plus State-of-Charge (SoC) and State-of-Health (SoH) Alerting and thus maintenance management support for Electric Vehicle (EV) batteries.
Our approach to the synthesis of high-fidelity EV battery-user-stories and automated battery lifecycle datasets generation for EV batteries, follows our methodologically-guided (UI-REF) framework. This provides for an ontologically committed use-context-specific modelling of battery life at two inter-related levels:
- the battery Electro-Chemical-and-Operational Specifications (EC-OS)- by Design;
- the Battery Lived Experience (BLE) – encompassing deployment (in-service) conditions of battery life, including: A) Charging Patterns (frequency and depth, DC fast charging vs AC slow charging -fast/slow – ,% charging completion), B) Daily load profile (highway, city, mixed, long-idle-periods), C) Peak-Load Bursts (towing, steep hills), D) Ambient Conditions (temperature, vibration levels) and E) the user’s battery maintenance regime
In this way, the synthetic battery life datasets generated through exhaustive permutation of all lifecycle use-context-permissible variations of both by-design and BLE life instances parametrics constitutes a powerful means of high-fidelity battery data generation to encompass not only the space of all known lifecycle conditions of possible battery life instances but also all hitherto unknown but conceivable future battery life instances – parametrics modelling to support our TinyML validation use-case.
References
[1] Topalli, N. and Badii, A. (2025) A user-centric context-aware framework for real-time optimisation of multimedia data privacy protection, and information retention within multimodal AI systems. Sensors, 25 (19). 6105. ISSN 1424-8220, doi: https://doi.org/10.3390/s25196105
[2] Badii et al (2009) Accessibility-by-Design: A Framework for Delivery-Context-Aware Personalised Media Content Re-purposing. In: Holzinger, A., Miesenberger, K. (eds) HCI and Usability for e-Inclusion. Lecture Notes in Computer Science, vol Springer, Berlin,Heidelberg (https://doi.org/10.1007/978-3-642-10308-7_14).
[3] Fredrickson, B. L., & Kahneman, D. (1993). Duration neglect in retrospective evaluations of affective episodes. Journal of Personality and Social Psychology, 65(1), 45–55. https://psycnet.apa.org/doiLanding?doi=10.1037%2F0022-3514.65.1.45
[4] Newell & Bröder (2008) Cognitive processes, models and metaphors in decision research. https://doi.org/10.1017/S1930297500002400
[5] Moin, A., Challenger, M., Badii, A. and Günnemann, S. (2022) Supporting AI engineering on the IoT edge through model-driven TinyML. In: 2022 IEEE 46th Annual Computers, Software, and Applications Conference (COMPSAC), 27 June 2022 – 01 July 2022, Los Alamitos, CA, pp. 884-893. doi: https://doi.org/10.1109/COMPSAC54236.2022.00140 (ISSN: 0730-3157, ISBN 978166548810-5).
